Search for: All records

Efficiently serving large language models (LLMs) under dynamic and bursty workloads remains a key challenge for real-world deployment. Existing serving frameworks and static model compression techniques fail to adapt to workload fluctuations, leading to either service-level objective (SLO) violations under full-precision serving or persistent accuracy degradation with static quantization. To deal with these issues, we present MorphServe, a dynamic, workload-aware LLM serving framework based on morphological adaptation. MorphServe introduces two asynchronous, token-level runtime mechanisms: quantized layer swapping, which selectively replaces less impactful layers with quantized alternatives during high-load periods, and pressure-aware KV cache resizing, which repurposes the freed memory to dynamically expand KV cache capacity. These mechanisms enable state-preserving transitions that jointly coordinate weight precision and KV capacity at runtime. Extensive experiments on Vicuna and Llama family models with real-world workloads demonstrate that MorphServe reduces average SLO violations by 92.45% and improves P95 TTFT by 2.2×–3.9× over full-precision serving, without compromising generation quality. Compared to planning-based quantization methods, MorphServe reduces average accuracy degradation by 41.3%, and lowers P95 TTFT by up to 2.4× over KV cache compression while maintaining higher generation quality. These results establish MorphServe as a practical and elastic solution that effectively navigates the accuracy–efficiency Pareto frontier under dynamic LLM serving workloads. 

Soft epidermal magnetic material has demonstrated excellent flexibility and wearability across a range of emerging applications, e.g., pressure sensing, robotics, among others. In this paper, we build a flexible magnetic skin based system MagSkin for touchless Human Computer Interaction (HCI). For example, a user can wear a Bandaid-like sticker on his fingertip to operate a smartphone using classic gestures without touching the screen, i.e., tap virtual icons floating on top of the screen. The magnetic skin is specially fabricated with randomly distributed magnetic microparticles, so that any shape change of the skin will cause position and orientation changes of those particles, introducing complex variations of the magnetic field around the skin. To characterize the magnetic field of the magnetic skin, we designed a multi-dipole based model. Based on this model, we developed 6-DoF position tracking and orientation estimation algorithms to localize the skin precisely. We also designed novel gestures, e.g., finger bend and pinch, for users to make input/adjustment based on the flexibility of the material. We have conducted experiments with 15 volunteers for 18 hours, experimental results show localization accuracy is 6.2 mm in 3D, 4.5 mm in 2D, and recognition accuracy is 94% for classical gesture and 92% for novel gesture. 

Multi-user environments present significant challenges in coordinating diverse preferences and resolving conflicts around shared resources. Current systems use a single-agent approach that struggles to balance individual needs with collective objectives. We introduce MALLM, a novel framework that deploys personalized LLM-based agents for each user on edge devices. MALLM integrates multi-sensor data fusion with a structured multi-agent decision-making mechanism, processing all data locally for enhanced privacy. Our edge-computing architecture enables real-time deliberation through evidence-based argumentation and consensus formation algorithms. The system continuously refines user profiles through sensor data while managing computational resources e!ciently. We evaluate MALLM through two case studies-health monitoring and personalized comfort management- demonstrating improved conflict resolution and resource e!ciency compared to conventional approaches. Our results show that MALLM e''ectively balances competing user priorities while preserving privacy in complex shared environments. 

Two-dimensional layered electrically conductive metal–organic frameworks (EC MOFs) have emerged as promising materials for electronic and energy applications due to their tunable electronic properties and structural versatility. Here, we employ machine learning (ML) models to predict the thermodynamic stability and electronic properties of EC MOFs, significantly reducing the dependence on costly ab initio calculations. We construct different feature sets by integrating compositional features from generic statistical reduction methods along with structural descriptors curated from the EC-MOF database (termed GD, M-GD, and A-GD features). Various ML models, including linear, tree-based, and ensemble learning approaches, are benchmarked for predicting formation energies of EC MOFs, as one measure of synthesizability, as well as predicting their electrically conductive nature. The density functional theory data pool spans the 536 monolayer systems in the EC-MOF database, divided into 90% training and 10% test sets. Our results demonstrate that ML models treated by proper feature engineering can achieve very high accuracy for formation energy prediction, with the coefficient of determination, R2, being as high as 0.96. Via proper feature engineering, we also report up to 92% accuracy in predicting the metallicity of EC MOFs and 82% in bandgap classification using the extra tree classifier. The trained ML models are further applied to a new class of EC MOFs, which are neither part of the training nor members of the EC-MOF database, showcasing their predictive power and transferability. This work establishes the foundations of a data-driven framework for accelerating the discovery of novel EC MOFs in future research. 

Abstract We study the asymptotics of the point process induced by an interacting particle system with mean-field drift interaction. Under suitable assumptions, we establish propagation of chaos for this point process: It has the same weak limit as the point process induced by i.i.d. copies of the solution of a limiting McKean–Vlasov equation. This weak limit is a Poisson point process whose intensity measure is related to classical extreme value distributions. In particular, this yields the limiting distribution of the normalized upper order statistics.